AI agent glossary
Agent trajectory
The full sequence of steps an agent took during a run, including every model turn, tool call, and result, which you can record and grade.
Two agent runs can end with the same answer while taking very different paths. One took three clean steps. The other took eleven, called a tool it was not allowed to use, and got lucky. The trajectory is the record of that path.
Evaluating trajectories means grading more than the final output: whether required steps happened, whether any safety rule was broken, and how many steps, tokens, and seconds the run used. Most checks are plain code over the recorded steps, sometimes with a model-based judge that reads the whole run.
Grade the outcome first and the path only where it matters, such as a required verification step or a forbidden action. Over-specifying the path penalizes agents for finding a valid route you did not anticipate.
A support agent's eval case requires that it look up the order before replying, never issue a refund above the limit, and finish in under eight tool calls.
Learn more in the course
Related terms
Go from definitions to doing
Learn to use and build AI agents in 130 practical lessons.